Text summarization is an incredibly useful tool. In journalism, a summary allows a busy reader to quickly grasp a story’s main points. In business, a summary of a meeting transcript allows an employee returning from vacation to catch up on what they missed. In customer service, a conversation summary helps an agent swiftly refresh on the contents and outcomes of previous interactions with a customer.
Historically, text summaries were composed by hand. However, given their broad usefulness, machine learning researchers have spent significant time and effort building algorithms that automate this process outright; in other words, a system that receives text as input—a news article, meeting lines, a customer service conversation, etc.—and outputs an accurate textual summary—instantly.
Broadly, text summarization techniques fall into two main camps: extractive and abstractive. In extractive summarization, the algorithm selects “salient” snippets of text then combines them into paragraph form. With this technique, summary contents are “accurate”—they are plucked verbatim from the original text—but the overall summary is often choppy and not intuitive. Conversely, in abstractive summarization, a novel summary paragraph is generated from scratch—as a human might do herself. In this setting, the overall summary is typically smooth and cohesive, yet there is an increased risk that its contents may be factually inconsistent. As such, both techniques imply a different set of tradeoffs, and each may be better suited to some applications more than others.
Taken together—countless use cases for summarization and significant research into the algorithms involved—numerous vendors have emerged to provide general-purpose summarization APIs. In this vein, they typically train their system on general-purpose datasets and, as per mainstream summarization research, produce summaries in paragraph form. Overall, these vendors often succeed in providing quality summaries for a broad range of consumers. However, it’s less clear that the algorithms, data, and output format they use will reliably produce excellent results for CX needs.
At ASAPP, we don’t build products for everyone. Instead, we focus singularly on building world-class technologies for customer experience teams. On this note, we set out to create a conversation-summarization system tailored specifically to the realities and operational needs of the CX domain.
Broadly, we believe that such a system should have the following key characteristics:
- The system should be accurate, via training on large amounts of contact center (both for service and sales) conversations—specific to the language of the business.
- The system should be scalable—it should be fast to deploy to new customers.
- Summary contents should be aggregatable—a supervisor should be able to easily build an analytics dashboard visualizing counts, trends, anomalies and more. For example, it should be simple to count how many times a given plan was offered and accepted, a given up-sell effort was successful, etc. (With a summary written in paragraph form, this is not straightforward to do!)
And indeed, we’ve built just that.
Generic summarization providers offer quality, free-form summaries to a broad customer base. ASAPP offers summaries built from aggregatable CX concepts trained on CX data. In other words, accurate and scalable summarization built specifically for you.
Crucial to our summarization system are three simple yet key competitive advantages—capabilities that ASAPP has, that others might not.
First, we make use of a single ontology of concepts that, in sum, broadly capture the possible content of customer service (i.e. contact centers that support care and sales calls) conversations across verticals. Internally, we call this ontology our Unified Meaning Representation (UMR). Externally, we’ve published (an abridged version of) our findings as Action-Based Conversations Dataset in NAACL 2021—offering a new task-oriented dialog dataset to the machine learning community at large. In the example above—”count how many times a given plan was offered and accepted”— comes directly from this ontology. As such, counting its occurrences is straightforward. In addition, given UMR’s design, we don’t have to create a new ontology for each new customer company.
With UMR, we’ve built a system that gives aggregatable content, which can be deployed in days.
Second, we train the system on large volumes of historical customer service conversations specific to each of our customer companies. In other words, your system is trained on your data—learning your unique lexicon (on an ongoing basis). In this vein, we drive accuracy specifically in your domain by teaching it the language of your agents and customers.
Finally, we rely on our in-house machine learning expertise to build the right models to tie the UMR ontology, and these data, together—creating a summarization system that does what you need it to do.
We have two options for managing conversation summaries.
Once a conversation has finished, we suggest UMR tags to the agent to use as part of their summary (with the option to add their own notes as they see fit).
Alternatively, our system can compose structured summaries—clearly denoting “Contact Reason,” “Actions Taken,” “Resolution,” etc.—in a fully-automatic manner—assisting the agent even further by enabling them to immediately jump to the next waiting customer.
In both cases, our system is taught customer service specific concepts and trained on customer service conversations. It is accurate, scalable, and offers summaries with aggregatable contents enabling downstream analytics. In other words, it was designed with you in mind from the start.
Interested? Send us a note at email@example.com and we’ll be happy to share more about how our AI-driven conversation summarization works and what it can do for you.
Will Wolf is a Staff Machine Learning Engineer at ASAPP, where he has researched, developed and deployed systems for: intent classification, personalized text recommendation and conversation summarization. In addition, he has researched and developed methods for dialog generation, dialog segmentation, “procedure induction” in goal-oriented dialog and online learning. Will holds a B.Sc. in Industrial Engineering from the Schreyer Honors College at Penn State University. In his free time, he likes to ride his bicycle and study foreign languages.